Moisture insensitive analysis of polyester/viscose waste textiles using Near-Infrared spectroscopy and Orthogonalization of external parameters algorithm

Abstract

Near-Infrared (NIR) spectroscopic analyses can be applied in waste textile recycling as a rapid and non-invasive method to provide both qualitative and quantitative results. However, it has been a challenge to enhance the accuracy rate of NIR-based waste textile sorting due to the major influences from water contexts in the samples. Orthogonalization of External Parameters (EPO) has been introduced to reduce the interference from water absorption in NIR spectral signals for better accuracy and reliability in modeling. Here we explore the feasibility of applying EPO strategy with varieties of algorithms, including partial least squares regression (PLS), artificial neural network (ANN), decision tree (DT), random forest (RF), gradient boosting decision tree (GBDT), extreme random tree (Extra-tree), decision tree model based on AdaBoost algorithm (AdaBoost-tree), support Vector machine (SVM), one-dimensional convolutional neural network (1D-CNN), and one-dimensional convolutional neural network with improved Inception structure (1D-Inception-CNN). 216 waste textiles samples from Xinjiang, China, were studied with different moisture levels. Among them, 80 samples were used to develop the EPO algorithm, 112 were used to establish the prediction models, and 24 were used as test datasets. Then, the samples were scanned using a near-infrared spectrometer at different moisture regain rates. Our results showed that the moisture content of waste textiles had strong absorption peaks near 1150 and 1450 nm, leading to a decrease in the near-infrared reflectance of waste textiles. To verify the effectiveness of the EPO algorithm, the decision coefficients (R² score) and other indicators of the model without the EPO process and the model with EPO process are systematically compared. Our results show that the EPO algorithm preprocessing improves the accuracy of the NIR model (The average decision coefficient (R² score) of the models was increased by 0.83), especially when the moisture interference level is significant. Therefore, the EPO integrated modeling method is a reliable approach for better accuracy in NIR-based waste textile sorting.

Keywords

Recycling of waste textiles near-infrared spectroscopic sorting of waste textiles deep learning machine learning

Introduction

Solid waste recycling has become one of the major constraints to sustainable development due to resource scarcity worldwide.¹ Since the 1980s, global textile fiber production’s average annual growth rate has been about 3%, totaling 107 million tons in 2018. Polyester and cotton are the most produced fibers, with projected production of 55 million tons and 26 million tons, respectively. Apparel accounts for 60% of global textile consumption, with a recent estimate of annual fiber production of post-consumer clothing at 53 million tons, of which only 12% is recycled, and 73% ends up in landfills or waste incinerators.² In 2019, the world’s demand for fiber exceeded 100 million tons, expected to increase to 121 million tons by 2025. The solid waste generated by the textile and garment industry accumulates yearly, hindering the development of the textile and garment industry and harming the ecological resources and environment.³

Due to the advantages of near-infrared spectroscopy, which avoids sample preparation, allows non-destructive, and provides rapid analyses in real-time, recent studies show that near-infrared spectroscopy (NIRS) has great potential for waste textile analysis.⁴ but the impact of external factors, such as moisture regain, can significantly influence the accuracy of quantification models.

Kirsti Cura et al. summarized the challenges faced by using near-infrared spectroscopy for sorting waste textiles, as follows:⁴ (1). Effects of Coatings and Finishes. (2). Effect of Blends/Elastane. (3). Effects of Structures. (4). Effects of Ageing on Cotton Fabrics. (5). Effects of Mercerisation on Cotton Fabrics. (6). Effect of Colours. However, they did not discuss the impact of moisture regain on analyzing waste textiles using near-infrared spectroscopy. Waste textiles usually contain water (also known as the moisture regains of waste textiles), which has strong absorption bands in the NIR region that may interfere with the spectral signature of other components.⁵ The hydration water has two strong OH absorption characteristics near 1400 and 1900 nm in the NIR region. When there is hygroscopic or adsorbed water on the surface area of a substance, it typically has an absorption band of about 2200 nm. Free water has strong absorption characteristics near 1400 and 1900 nm, and weak absorption peaks near 980 and 1200 nm. It can also affect measurement results by reducing the reflectivity of the material.⁶ Therefore, we believe that the near-infrared spectral data of waste textiles are not only affected by their composition, but also by their moisture regain.

The research on the influence of moisture on the determination of soil composition using near-infrared spectroscopy has been relatively mature. Budiman Minasny et al. used the external parameter Orthogonalization (EPO) algorithm to remove the impact of soil moisture in the NIR spectrum to calibrate the SOC content.⁷ Their work has proven that the EPO algorithm can effectively remove changes in soil moisture. In addition, Real-time soil moisture content (SMC) monitoring is an important parameter in precision agriculture that can be utilized to enhance soil and water management. Visible and near-infrared (vis-NIR) has been proposed as a promising method for SMC monitoring. However, vis-NIR reflectance response to soil moisture is strongly influenced by soil properties such as texture and organic matter content. Thus it is difficult to develop a general prediction model of vis-NIR that can estimate SMC of different soil types. The results of Jiang Liu et al. indicate that the EPO algorithm can achieve a generalized SMC prediction model.⁸ The EPO algorithm can also remove the influence of temperature on near-infrared spectroscopy. Ren Sheng et al. used mixed temperature correction method and external parameter orthogonal method (EPO) to reduce the impact of temperature changes on the NIR spectra of soluble solids and Lycopene in Cherry tomato. Their experimental results show that the EPO method has better prediction results than the mixed temperature correction model.⁹

In this study, we investigate the impact of moisture regain on the NIRS of polyester/viscose blended waste textiles and evaluate the effectiveness of the EPO algorithm in correcting moisture-induced variations in the NIRS. We collected fabric samples of polyester/viscose blended waste textiles from Xinjiang Ruyi Textile and Clothing Co., Ltd.¹⁰ The experiment was designed and implemented to humidify waste textiles artificially blended with polyester/viscose fibers. A total of 7 moisture levels were obtained Near-infrared spectral data of waste textiles blended with polyester/viscose fibers. The characteristics of the impact of moisture regain on the Near-infrared spectral absorption of waste textiles blended with polyester/viscose fibers were analyzed, and the effects of the EPO algorithm for removing moisture on waste textiles blended with polyester/viscose fibers were emphatically studied. Using partial least squares regression (PLS), artificial neural network (ANN), decision tree (DT), random forest (RF), gradient boosting tree (GBDT), extreme random tree (Extra-tree), decision tree model based on AdaBoost algorithm (AdaBoost-tree), support Vector machine (SVM), one-dimensional convolutional neural network (1D-CNN), and one-dimensional convolutional neural network with improved Inception structure (1D-Inception-CNN) to establish a quantitative model of waste textiles before and after EPO treatment. Then, the effectiveness of the EPO algorithm implementation is evaluated by comparing the determination coefficient (R²) and root mean square error (RMSE) of the model before and after EPO processing.

To the best of our knowledge, this is the first study to investigate the effectiveness of the EPO algorithm for correcting the impact of moisture regain on the NIRS of waste textiles. The results of this study have the potential to improve the accuracy and reliability of NIRS-based models for waste textile sorting and recycling. By demonstrating the effectiveness of EPO in correcting the impact of external factors, this study contributes to the larger effort to promote sustainable waste management.

Material and methods

Textile samples and dataset description

The fabric samples of waste textiles used in this study were obtained from Xinjiang Ruyi Textile and Clothing Co., Ltd. The manufacturer has labeled each fabric sample with ingredient information. The fabric samples are all blended with polyester and viscose fibers. Generally, Near-infrared spectral sequence data is defined as:

X = (\begin{array}{c} x_{1} & x_{2} & \dots & x_{m} \end{array})

(1)

Where,

X

is the near-infrared spectral data sequence matrix corresponding to a fabric sample,

m

is the number of wavelength points,

x_{m}

is the corresponding absorption coefficient. When the

X

matrix contains

n

Near-infrared spectral sequence data of fabric sample, can be expressed as:

X = (\begin{array}{c} x_{11} & \dots & x_{1 m} \\ ⋮ & ⋱ & ⋮ \\ x_{n 1} & \dots & x_{n m} \end{array})

(2)

Therefore, the fabric sample spectral dataset is defined as:

S = {X_{1}, X_{2} \dots X_{c}}

(3)

Where,

S

is the Near-infrared spectral data set of fabric samples,

c

is the spectral data matrix number.

Due to the complexity of the source and composition of waste textiles, and the vulnerability of the moisture regain of waste textiles to humidity and other environmental factors, the moisture regain of waste textiles is not fixed, making it difficult to establish accurate near-infrared spectroscopy prediction models for qualitative analysis of waste textiles. Therefore, the purpose of this study is to achieve moisture-insensitive prediction of waste textiles, and three data sets were designed for this purpose:

Spectral dataset for dry fabrics samples( $S_{m i x}$ ): This data set was obtained by scanning 112 fabric samples. Each sample of waste textiles was scanned seven times with different moisture regain. This group of fabric samples was scanned once with a moisture regain of 0% and scanned six times at different moisture regain levels.

Spectral dataset of fabric samples for EPO algorithm development( $S_{e p o}$ ): This dataset was scanned from 80 fabric samples and used to establish an EPO model(To obtain the matrix). Each sample of waste textiles was scanned seven times with different moisture regain. This group of fabric samples was scanned once with a moisture regain of 0% and scanned six times at different moisture regain levels. The following section will describe the wetting experiment of fabric samples in detail.

Spectral dataset of fabric samples for testing( $S_{t e s t}$ ): This dataset comprises 24 fabric samples for model validation. Under the same humidity conditions as the EPO development dataset, samples in this group were also scanned 6 times (no scanning was performed when the moisture regain was 0%).

The statistical characteristics of each dataset are shown in Table 1. The dataset contains waste textiles in various colors, but the components are all polyester and viscose fibers. The reason for the three data sets

S_{m i x}

S_{e p o}

, and

S_{t e s t}

is to verify the effectiveness of the EPO algorithm. The data modeling process for verifying the effectiveness of the algorithm is shown in Figure 1. (In Figure 1, the mathematical symbols of the dataset processed by the EPO algorithm are

S_{m i x}^{*}

and

S_{e p o}^{*}

Table 1.

Statistical table for three datasets.

DataSet	Color	Polyester fiber content/%	Viscose fiber content/%	Moisture regain/%				Number of fabric sample	Number of spectral data
DataSet	Color	Polyester fiber content/%	Viscose fiber content/%	Mean	Median	Max	min	Number of fabric sample	Number of spectral data
S_epo	Khaki	100	0	15	15	30	0	4	28
	Grey	70	30	15	15	30	0	4	28
	Grey	80	20	15	15	30	0	4	28
	Red	70	30	15	15	30	0	4	28
	Tibetan blue	65	35	15	15	30	0	12	84
	Tibetan blue	80	20	15	15	30	0	8	56
	Wine red	100	0	15	15	30	0	12	84
	Wine red	100	0	15	15	30	0	4	28
	Sallow	70	30	15	15	30	0	8	56
	Black	65	35	15	15	30	0	8	56
		70	30	15	15	30	0	4	28
		80	20	15	15	30	0	8	56
S_mix	Khaki	100	0	15	15	30	0	6	42
	Grey	70	30	15	15	30	0	5	35
	Grey	80	20	15	15	30	0	5	35
	Red	70	30	15	15	30	0	5	35
	Tibetan blue	65	35	15	15	30	0	18	126
		80	20	15	15	30	0	13	91
		100	0	15	15	30	0	16	112
	Wine red	100	0	15	15	30	0	5	35
	Sallow	70	30	15	15	30	0	11	77
	Black	65	35	15	15	30	0	12	84
		70	30	15	15	30	0	6	42
		80	20	15	15	30	0	10	70
S_test	Red	70	30	17.5	17.5	30	5	3	18
	Tibetan blue	65	35	17.5	17.5	30	5	3	18
		80	20	17.5	17.5	30	5	4	24
		100	0	17.5	17.5	30	5	5	30
	Wine red	100	0	17.5	17.5	30	5	1	6
	Sallow	70	30	17.5	17.5	30	5	1	6
	Black	65	35	17.5	17.5	30	5	3	18
		70	30	17.5	17.5	30	5	2	12
		80	20	17.5	17.5	30	5	2	12

Figure 1.

Flow chart of external parameter orthogonalization (EPO) development and validation scheme.

Fabric sample drying and NIR scanning

To achieve a moisture regain of 0% for fabric samples, the experiment used a Y802 N dryer (from Changzhou First Textile Equipment Co., Ltd.) to dry the fabric samples. The fabric samples underwent continuous drying at 65°C for 120 minutes to remove all moisture components.

After drying, the fabric samples are instantly analyzed by a DA200 universal online near-infrared spectrometer (from Tianjin Jiuguang Technology Development Co., Ltd.) in the air atmosphere with a temperature of 20°C and humidity of 50%–60% to collect spectral data of dried fabric samples one by one in the darkroom. The DA200 Near-infrared spectrometer has a spectral range of 950 nm to 1650 nm and a resolution of 5 nm. The upper software of the spectrometer is also developed by Tianjin Jiuguang Technology Development Co., Ltd. The algorithm and model of this study were written in Python.

Each fabric sample was scanned four times on both the front and back sides to calculate the average spectrum as the overall spectrum of the fabric sample. If the fabric sample is too thin, it is folded into 2 or 3 layers before scanning.

Rewetting procedure and NIR scanning

Seven polyester/viscose waste textile fabric samples (0%, 5%, 10%, 15%, 20%, 25%, 30%) were prepared to investigate the effect of moisture regain changes on near-infrared spectroscopy. Each fabric sample is dried (with a moisture content of 0%), quickly weighed, and weighed its dried fabric sample (dry weight of the fabric). Then, quickly place the sample in a rectangular black box with a side length of 20 cm and a depth of 5 cm. We used a spray to evenly spray water on the fabric sample to obtain an incremental moisture level of 5% per sample, and then quickly sealed the sample to prevent water evaporation. The weight of the fabric sample to be sprayed with water is calculated as follows:

w_{w a t e r} = (1 + r_{r e g a i n}) \times w_{d r y T e x t i l e} - w_{d r y T e x t i l e}

(4)

Where,

w_{w a t e r}

is the weight of polyester/viscose waste textile samples to be sprayed with water.

r_{r e g a i n}

is moisture regain,

w_{d r y T e x t i l e}

is the dry weight of a sample of polyester/viscose waste textiles.

Then, the samples were allowed to balance and stabilize for 2h in a black box to fully absorb and distribute water evenly. Then the spectral data of each wet fabric sample was performed. Each fabric sample is scanned four times (two scans on the front and back sides). For thinner fabric samples, they were folded into 2 or 3 layers before scanning. The average spectrum of the 4 scans was adopted as the representative spectrum of the fabric sample. Next, each sample was immediately reweighed to calculate the actual moisture regain level. The water addition step was repeated until the experiment is fully completed.

Spectral preprocessing

The collected near-infrared spectral data of polyester/viscose waste textiles are subjected to standard normal variable (SNV) transformation pretreatment, by the following formula:

X_{(s n v)} = \frac{X - E [X]}{\sqrt{Var [X]}}

(5)

Where,

E [X]

is the row mean of the spectral sequence matrix,

\sqrt{Var [X]}

is the row standard deviation of the spectral sequence matrix.

Model calibration/validation and EPO transformation with model coupled CV and Wilk’s Λ

In Figure 2, We use the spectra from the $S_{e p o}$ set to develop the EPO algorithm. Firstly, divide the spectrum in the set into two parts: (1). $X_{D r y S e t}$ is the spectrum obtained from the fabric samples with a moisture regain of 0%. (2). $X_{M o s i t r u e S e t}$ is the spectrum obtained when the moisture regains of the fabric samples is not 0%. Grouping $X_{D r y S e t}, X_{M o s i t r u e S e t}$ by moisture regain and calculate the average value to $X_{d r y}$ and $X_{M o s i t r u e}$ . The EPO algorithm can solve for the $P$ matrix based on $X_{d r y}$ , $X_{M o s i t r u e}$ . $X_{C o r r e c t S e t}$ is the spectral matrix of $S_{e p o}$ after correction by the EPO algorithm.

Figure 2.

Flow chart of using EPO algorithm.

The principle of EPO was orthogonally projecting the original spectra to the subspace of the spectra influenced by external factors. Thus, the effects caused by external factors can be removed.

The EPO algorithm projects the spectrum of all polyester/viscose waste textile fabric samples onto a space orthogonal, allowing the removal of the moisture inferences to the spectral data.¹¹ The EPO algorithm assumes that the original spectral matrix consists of two parts: a useful component, and a parasitic component caused by external factors (Moisture in this case). The mathematical assumption formula is as follows:

X = X P + X Q + R

(6)

Where,

X P

is a useful component,

X Q

is a parasitic component,

R

is residuals matrix. In this study,

X P

is the spectral portion caused by the fabric sample,

X Q

is a parasitic part of the spectrum caused by moisture. The most critical part of the EPO algorithm is the solution of the

P

matrix. The common method for solving the

P

matrix is as follows: First, solve the difference

D

matrix as follows:

D = X_{M o s i t r u e} - X_{d r y}

(7)

Then, singular value decomposition is performed on the difference matrix

D

[U, S, V] = SVD (D^{T} D)

(8)

Next, according to the super parameter

n p c

defined by the EPO algorithm, select a subset from

V

V_{s} = V {[0 : n p c, :]}^{T}

(9)

Q can be solved as

Q = V_{S} {V_{S}}^{T}

(10)

According to the above formula, it can be deduced that:

P = I - Q

(11)

Where

I

is an identity matrix. To facilitate understanding of this algorithm, the pseudocode of the algorithm is shown in Figure 3. A hyperparameter in the pseudocode needs to be determined, which is very important for the modeling of EPO algorithm. There are two methods for determining hyperparameters

n p c

.(1) Change the hyperparameter

n p c

of the EPO algorithm and perform cross-validation with the model to select the best

n p c

. (2) Calculate Wilk’s Λ of transformation spectrum

X

W i l k ’ s Λ = \frac{Trace (B)}{Trace (T)}

(12)

Where,

T

is the variance-covariance matrix of the

X

spectral matrix after EPO transformation (The variance-covariance matrix of

S_{m i x}^{*}

in the text).

B

is the variance-covariance matrix of the average spectrum of each moisture gradient in the

X

spectral matrix after EPO transformation (In the text, it is the variance-covariance matrix of the average spectrum of each moistrue levels of

S_{m i x}^{*}

Figure 3.

Pseudocode of EPO algorithm.

Model construction and evaluation indicators

After obtaining a dataset of near-infrared spectra of waste textiles, a quantitative model of near-infrared spectra of waste textiles before and after EPO processing was constructed using common machine learning and deep learning algorithms. Finally, the decision coefficient (R²) and error of the model prediction are evaluated to determine the effectiveness of the EPO algorithm. Next, we will introduce the corresponding models.

Decision Tree (DT) regression is a commonly used and effective supervised learning method using a top-down recursive strategy.¹² Decision tree (DT) is easy to understand and interpret, and can handle nonlinear relationships between input features and output variables. Adaboost-tree Regression is an integrated learning method based on the Adaboost algorithm. It combines the advantages of Decision Tree (DT) regression and the Adaboost algorithm, and can effectively handle regression problems. However, adaboost-tree is easy to overfit. Random Forest (RF) is also an integrated learning method based on decision trees(DT), which uses bagging techniques to create new training sets. It includes two essential methods: random feature subspace and out-of-package estimation. The former can construct the tree faster, while the latter can assess the relative importance of each input feature. Random forest (RF) has a stronger generalization ability and is not easy to overfit compared to partial least squares (PLS).¹³ Extreme Random Tree (Etrra-tree) Regression is also an integrated learning algorithm based on decision trees, similar to Random Forest (RF), but in the training process of Decision Tree (DT), Extra-Tree randomly selects partition points, further increasing the randomness of the model. Gradient boosting decision tree (GBDT) during the training process of each weak regressor, the gradient boosting decision tree (GBDT) will be weighted based on the residuals of the previous round, thereby gradually improving the fitting effect of the model. Gradient boosting decision tree (GBDT) regression models have high accuracy and robustness and can process datasets with high-dimensional features and large amounts of noise.

Support Vector Machine (SVM) is a learning algorithm with good generalization performance. It can model complex non-linear boundaries using adaptive kernel functions.SVM can handle high-dimensional data, can handle non-linear relationships between input features and output variables, finds the hyperplane that best separates the different classes of data points. Therefore,SVM has achieved excellent results in the classification and regression tasks of Near-Infrared (NIR) spectroscopy.¹⁴ However, It is also computationally expensive and challenging to interpret.

We also built a one-dimensional convolutional neural network (1D-CNN) with two one-dimensional convolutional layers. The structure of 1D-CNN is shown in Figure 4. A batch normalization layer is set behind each convolution layer of 1D-CNN to accelerate the training of neural networks and improve the accuracy of the model. Using the ReLU function as the activation function for 1D-CNN, set a Dropout layer on the last layer to enhance the generalization ability of the model. Table 2

Figure 4.

Schematic diagram of one-dimensional convolution network structure.

Table 2.

Architectural details of 1D-CNN and Improved 1D-Inception-CNN.

Layer	1D-CNN	Improved 1D-Inception-CNN
Input	1 × 141	1 × 141
Convolution/inception	C:1 × 3 × 64	C:1 × 3 × 8	C:1 × 3 × 8
Convolution/inception	C:1 × 3 × 64	C:1 × 3 × 8	C:1 × 3 × 8
Normalization	BN	BN
Activation	ReLU	ReLU
Concatenation		Concat
MaxPooling	1 × 3(strides = 2)	1 × 3(strides = 2)
Convolution/inception	C:1 × 3 × 64	C:1 × 3 × 16	C:1 × 3 × 16
Convolution/inception	C:1 × 3 × 64	C:1 × 3 × 16	C:1 × 3 × 16
Normalization	BN	BN
Activation	ReLU	ReLU
Concatenation		Concat
GlobalAveragePooling	128	32
FC	1	1
Output	Linear	Linear

Based on 1D-CNN, an improved 1D-Inception-CNN based on the Inception structure is proposed. Compared with 1D-CNN, the improved 1D-Inception-CNN has a wider receptive field and better model accuracy.¹⁵ The structure of the one-dimensional convolutional neural network based on the improved Inception we built is shown in Figure 5. This neural network uses two layers of improved Inception, each layer of convolution is also set with a batch normalization layer, and the ReLU function is used as the activation function of 1D-Inception-CNN.

Figure 5.

One-dimensional convolutional neural network with improved Inception structure.

One-dimensional convolutional neural network (1D-CNN) models have several advantages over traditional machine learning methods for modeling near-infrared spectroscopy (NIRS) data:1D-CNN models can automatically extract relevant features from NIRS data, without the need for manual feature engineering. And, 1D-CNN models are capable of learning complex patterns and relationships between spectral features, which can improve the accuracy of NIRS data modeling. Moreover, 1D-CNN models can be trained end-to-end, which can lead to faster and more efficient training.

However, there are also some disadvantages to using 1D-CNN models for NIRS data modeling:1D-CNN models may require large amounts of training data to achieve optimal performance.Interpretability of 1D-CNN models can be challenging, as they are often considered as black boxes.1D-CNN models can be computationally intensive, requiring powerful computing resources for training and inference.

The model performance was evaluated using the following indicators: determination coefficient (R²), root mean square error (RMSE), mean absolute error (MAE), logarithmic mean square error (MSLE), mean absolute percentage error (MAPE), median absolute error (MedAE), and maximum error (Max Error).

R^{2} (y, \hat{y}) = 1 - \frac{\sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})}{\sum_{i = 1}^{n} (y_{i} - {\bar{y}}_{i})}

(13)

RMSE (y, \hat{y}) = \sqrt{\frac{1}{n} {\sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})}^{2}}

(14)

MSLE (y, \hat{y}) = \frac{1}{n} {\sum_{i = 1}^{n - 1} (\log_{e} (1 + y_{i}) - \log_{e} (1 + {\hat{y}}_{i}))}^{2}

(15)

MAE (y, \hat{y}) = \frac{1}{n} \sum_{i = 1}^{n} | y_{i} - {\hat{y}}_{i} |

(16)

MAPE (y, \hat{y}) = \frac{1}{n} \sum_{i = 0}^{n - 1} \frac{| y_{i} - {\hat{y}}_{i} |}{\max (ε, | y_{i} |)}

(17)

MedAE (y, \hat{y}) = {median (| y}_{1} - {\hat{y}}_{1} |, \dots {, | y}_{n} - {\hat{y}}_{n} |)

(18)

MaxError (y, \hat{y}) = \max | y_{i} - {\hat{y}}_{i} |

(19)

Where,

y

is the actual value.

\hat{y}

is the value of model estimate.

Results and discussion

Analysis of the spectral characteristics of fabric samples with different moisture content levels and the difference spectrum between wet and dry fabric samples

The average spectra of fabric samples for each moisture levels are shown in Figure 6(a). It is clear that the absorbance of the fabric samples increases with the moistrue levels of fabric sample. The absorption peak at 1450 nm is the most prominent. When textiles are exposed to liquids or dry textiles are exposed to moist air, different areas of them may be wetted: the liquid may be between fibers, between yarns, or in the film and gaps of fibers; Or inside fibers, inside fiber pores, inside hollow fiber cavities (such as cellulose or cotton), or inside materials. Due to the fact that each fabric sample with the same moisture level is a Polyester/Viscose textile.¹⁶ So, the main factor that affects the spectrum is the free water inside waste textiles.

Figure 6.

Average spectrum. (a) The average spectrum of fabric for each moisture levels. (b) Average spectrum after SNV transformation.

The average spectra of various moisture levels after SNV preprocessing are shown in Figure 6(b). The differences caused by the morphology of the fabric samples between various moisture levels were slightly reduced, but there were still apparant differences in the absorption peak shape at 1150 and 1450 nm. This indicates that although the SNV algorithm has a significant effect on reducing scattered light on the surface of fabric samples and removing background noise interference, it cannot remove the influence of the moisture regain rate of waste textiles in near-infrared spectra. The phenomenon of deviation in the spectral curve of the fabric caused by the moisture reduces the accuracy of qualitative models.

Determination of hyperparameters in EPO algorithm

To find the optimal hyperparameter of the EPO algorithm. We solved the Wilk’s Λ with $n p c$ from 1 to 10, the results are shown in Figure 7. When $n p c$ is 2, Wilk’s Λ reaches the maximum value. The optimal hyperparameter of the EPO algorithm should be around 2. Then, the EPO algorithm with different $n p c$ hyperparameters was used to correct the training and test sets to obtain the corrected training set ( $S_{m i x}^{*}$ ) and test set ( $S_{t e s t}^{*}$ ). The data of the corrected training set is used to establish a quantitative model, and the corrected test set is used to test the model. The results are shown in Figure 8. The PLS and Extra-Tree models achieve the best decision coefficient (R2) when $n p c$ is 2, and the other models achieve the best determination coefficient (R2) when $n p c$ is 3. Both 1D-CNN and 1D-Inception-CNN models work best when $n p c$ is 1. This result is consistent with the results obtained in other literature. Basically, 2 or 3 is the optimal hyperparameter for the EPO algorithm.¹⁷

Figure 7.

The image of Wilk’s ∧ varies with the external parameter orthogonalization hyperparameter. A higher Wilk’s ∧ indicates that different fabric samples have a higher degree of separation in spectral space relative to the same sample with different moisture levels.

Figure 8.

Cross-validation results. (a) Changes in R² scores of regression models for various machine learning EPO algorithms with different hyperparameters after data preprocessing. (b) Changes in R² scores of regression models for various machine learning EPO algorithms with different hyperparameters after data preprocessing(deep learning).

Analysis of waste textile spectral characteristics corrected by EPO algorithm

The visualization results after EPO algorithm processing are shown in Figure 9 and 10. The average spectra of the fabric samples for each moisture level after EPO algorithm processing are close to consistency. Especially when the $n p c$ is 2, the average spectra of the fabric samples with different moisture levels is very close, indicating that the EPO algorithm can successfully separate useful information and parasitic components.By observing the $Q$ matrix, it can be found that the main influence on moisture is the absorption of the spectrum at 1450 nm and 1150 nm. Further indicating that the deviation of the spectral curve is mainly caused by the free water inside the textile.

Figure 9.

Visualization of an external parameter orthogonalization algorithm when the hyperparameter $n p c$ is 1 (a) Matrix $Q$ (b) Matrix $P$ (c) Parasitic part (d) Useful part.

Figure 10.

Visualization of an external parameter orthogonalization algorithm when the hyperparameter $n p c$ is 2 (a) Matrix $Q$ (b) Matrix $P$ (c) Parasitic part (d) Useful part.

Model evaluation

We trained machine learning and deep learning models using the content of viscose fibers as labels. Table 3 displays the performance of common machine learning models, which show that all pre-EPO machine learning models had significant error due to the spectral curve offset caused by moisture in waste textiles. Post-EPO algorithms greatly reduced prediction error and improved the R2 score. Prior to EPO algorithm processing, the average Coefficient of Determination score for the eight machine learning algorithms was only −1.955, which significantly improved to 0.83 after EPO algorithm processing. In previous studies on the effect of soil moisture on near-infrared spectroscopy. Without EPO, most of the machine learning models' Coefficient of determination scores are between 0.4 and 0.6,¹⁸ which proves that the influence of fabric moisture on the near-infrared spectrum is stronger.

Table 3.

Machine learning model scoring table before and after EPO algorithm processing.

Models	Processing method	R2 score	RMSE	MAE	MLSE	MAPE	MedAE	Max error
ANN	Not corrected by EPO	−1.84	0.20	0.16	0.01	0.21	0.13	0.49
ANN	Corrected by EPO	0.87	0.05	0.04	0.00	0.05	0.03	0.12
Decision tree	Not corrected by EPO	−1.41	0.20	0.16	0.01	0.20	0.13	0.35
Decision tree	Corrected by EPO	0.81	0.06	0.02	0.00	0.02	0.00	0.30
PLS	Not corrected by EPO	−1.84	0.20	0.16	0.01	0.21	0.13	0.50
PLS	Corrected by EPO	0.87	0.05	0.04	0.00	0.05	0.03	0.12
SVM	Not corrected by EPO	−2.53	0.19	0.15	0.01	0.21	0.13	0.44
SVM	Corrected by EPO	0.72	0.06	0.05	0.00	0.06	0.04	0.33
RandomFrest	Not corrected by EPO	−1.88	0.20	0.16	0.01	0.21	0.13	0.35
RandomFrest	Corrected by EPO	0.90	0.04	0.02	0.00	0.02	0.00	0.20
Adaboost	Not corrected by EPO	−2.32	0.18	0.14	0.01	0.18	0.10	0.35
Adaboost	Corrected by EPO	0.80	0.06	0.03	0.00	0.04	0.02	0.30
GBRT	Not corrected by EPO	−2.01	0.20	0.16	0.01	0.21	0.13	0.36
GBRT	Corrected by EPO	0.82	0.06	0.02	0.00	0.03	0.00	0.30
ExtraTree	Not corrected by EPO	−1.81	0.20	0.16	0.01	0.21	0.12	0.35
ExtraTree	Corrected by EPO	0.88	0.05	0.02	0.00	0.02	0.00	0.23

The mean absolute error (MAE) is used as the loss function in the training of 1D-CNN and 1D-Inception-CNN. The $S_{m i x}$ and $S_{t e s t}$ before and after processing with the EPO algorithm are used as training sets and validation sets for the deep learning model to train the 1D-CNN and 1D-Inception-CNN models, respectively. The change of loss value and R² score with Epochs is shown in Figures 11 and 12. It is clear that both 1D-CNN and 1D-Inception-CNN have serious overfitting problem before using the EPO algorithm to process $S_{m i x}$ and $S_{t e s t}$ . With the gradual fitting of 1D-CNN and 1D-Inception-CNN to $S_{m i x}$ , The error of 1D-CNN and 1D-Inception-CNN on the validation set $S_{t e s t}$ gradually rises. The performance is significantly enhanced in the case of post-EPO dataset, where overfitting problems are largely removed. Convolutional neural networks possess translation invariance, allowing for spectral analysis on raw material samples with varying batches or moisture levels. This effectively reduces noise and interference signals, making the EPO algorithm unnecessary in achieving good results. However, despite this potential, our experimental results show that neither 1D-CNN nor 1D-Inception-CNN could achieve satisfactory results without the EPO algorithm.

Figure 11.

The loss value and R² score change curve of the one-dimensional convolutional neural network before and after processing by EPO algorithm. (a) Loss value variation curve without using EPO algorithm. (b) R² variation curve without using EPO algorithm. (c) The loss value change curve using the EPO algorithm. (d) R² variation curve using EPO algorithm.

Figure 12.

The loss value and R² score change curve of the improved Inception convolutional neural network before and after the EPO algorithm processing. (a) Loss value variation curve without using EPO algorithm. (b) R² variation curve without using EPO algorithm. (c) The loss value change curve using the EPO algorithm. (d) R² variation curve using EPO algorithm.

According to the above experiments, it is easy to prove that the moisture of waste textiles will greatly impact on near-infrared spectroscopy, which will lead to extreme distortion of qualitative analysis of waste textiles by near-infrared spectroscopy. Using EPO algorithm to process the data can effectively improve the applicability and score of the model.

Conclusion

The analysis leads to the following conclusions:

1. Moisture significantly impacts the near-infrared spectra of waste textiles, and there is a strong absorption peak near 1150 and 1450 nm in the spectrum.

2. The overall spectral reflectance of waste textiles decreases and their absorption increases, as the moisture regained from them increases.The OH bonds in water molecules are the cause of this phenomenon, as found through relevant literature.

3. Using SNV as a preprocessing method is insufficient to correct the spectral measurement error caused by moisture, thus affecting the model’s analysis of waste textiles.

4. Both deep learning and machine learning are affected by the moisture content of waste textiles, resulting in low accuracy.

5. Adopting the EPO algorithm greatly improves the model’s accuracy and sorting capability of waste textiles. The progress is enormous.Near-infrared spectroscopy measurement of waste textiles in the laboratory can avoid the interference of many environmental and human factors, and the data obtained are of high quality and good repeatability. The accuracy of modeling improvement is more obvious. However, in the factory environment of waste textile sorting, the near-infrared spectrum of waste textiles is bound to be affected by color, weight, surface texture, fabric structure, environmental temperature, light conditions, and other possible factors. Since the moisture regains of waste textiles is the most essential part of the external environmental factors, we have demonstrated that EPO can be applied to reduce the influences from external noise.

We have verified the effect of fabric coating on the near-infrared spectrum of waste textiles, and the result is that fabric coating impacts the near-infrared spectrum of waste textiles. However, whether the EPO algorithm can also be used to correct this error requires further discussion. With the development of deep learning, using autoencoders to denoise data is a very effective method. Using an autoencoder to correct the near-infrared spectra of waste textiles affected by moisture is also a promising method, and further discussion is needed on its feasibility.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest to concerning the research, authorship, and or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Major Science and Technology Program of Xinjiang Autonomous Region: Development and Demonstration of Technology for Conversion and Utilization of Waste Textile Raw Material Resources(NO.2020A03002-4) and Project of Xinjiang Production and Construction Corps' Science and Technology Research Program: Technical Promotion of Intelligent Sorting and Packaging System for Cotton Spinning Yarn in Southern Xinjiang Textile Enterprises (NO.2019AB014).

ORCID iD

Xun Qiu

References

C-k

Cheng

Liao

, et al. An account of the textile waste policy in China (1991–2017). Journal of Cleaner Production 2019; 234: 1459–1470. DOI: 10.1016/j.jclepro.2019.06.283.

Mäkelä

Rissanen

Sixta

. Machine vision estimates the polyester content in recyclable waste textiles. Resources, Conservation and Recycling 2020; 161: 2010. DOI: 10.1016/j.resconrec.2020.105007.

Riba

Cantero

Riba-Mosoll

, et al. Post-Consumer Textile Waste Classification through Near-Infrared Spectroscopy, Using an Advanced Deep Learning Approach. Polymers (Basel) 2022; 14: 2022. DOI: 10.3390/polym14122475.

Cura

Rintala

Kamppuri

, et al. Textile Recognition and Sorting for Recycling at an Automated Line Using Near Infrared Spectroscopy. Recycling 2021; 6: 20. DOI: 10.3390/recycling6010011.

Knadel

Deng

Alinejadian

, et al. The Effects of Moisture Conditions-From Wet to Hyper dry-On Visible Near-Infrared Spectra of Danish Reference Soils. Soil Science Society of America Journal 2014; 78: 422–433. DOI: 10.2136/sssaj2012.0401.

Stenberg

Viscarra Rossel

Mouazen

, et al. Chapter Five - Visible and Near Infrared Spectroscopy in Soil Science. In: Sparks

(ed). Advances in Agronomy. Academic Press, 2010, pp. 163–215.

Minasny

McBratney

Bellon-Maurel

, et al. Removing the effect of soil moisture from NIR diffuse reflectance spectra for the prediction of soil organic carbon. Geoderma 2011; 167-168: 118–124. DOI: 10.1016/j.geoderma.2011.09.008.

Liu

Zhang

Yang

, et al. Developing a generalized vis-NIR prediction model of soil moisture content using external parameter orthogonalization to reduce the effect of soil type. Geoderma 2022; 1: 419. DOI: 10.1016/j.geoderma.2022.115877.

Sheng

Cheng

, et al. Model development for soluble solids and lycopene contents of cherry tomato at different temperatures using near-infrared spectroscopy. Postharvest Biology and Technology 2019; 156: 118. DOI: 10.1016/j.postharvbio.2019.110952.

10.

Roger

J-M

Chauchard

Bellon-Maurel

. EPO–PLS external parameter orthogonalisation of PLS application to temperature-independent measurement of sugar content of intact fruits. Chemometrics and Intelligent Laboratory Systems 2003; 66: 191–204. DOI: 10.1016/s0169-7439(03)00051-0.

11.

Liu

Deng

, et al. Evaluating the characteristics of soil vis-NIR spectra after the removal of moisture effect using external parameter orthogonalization. Geoderma 2020; 6: 376. DOI: 10.1016/j.geoderma.2020.114568.

12.

Ren

Wang

Ning

, et al. Using near-infrared hyperspectral imaging with multiple decision tree methods to delineate black tea quality. Spectrochim Acta A Mol Biomol Spectrosc 2020; 237: 118407. DOI: 10.1016/j.saa.2020.118407.

13.

Lee

Choi

Cha

, et al. Random forest as a potential multivariate method for near-infrared (NIR) spectroscopic analysis of complex mixture samples: Gasoline and naphtha. Microchemical Journal 2013; 110: 739–748. DOI: 10.1016/j.microc.2013.08.007.

14.

Devos

Ruckebusch

Durand

, et al. Support vector machines (SVM) in near infrared (NIR) spectroscopy: Focus on parameters optimization and model interpretation. Chemometrics and Intelligent Laboratory Systems 2009; 96: 27–33. DOI: 10.1016/j.chemolab.2008.11.005.

15.

Chai

Zeng

Lin

, et al. Improved 1D convolutional neural network adapted to near-infrared spectroscopy for rapid discrimination of Anoectochilus roxburghii and its counterfeits. J Pharm Biomed Anal 2021; 199: 114035. DOI: 10.1016/j.jpba.2021.114035.

16.

Duprat

. Moisture in Textiles. Annual Review of Fluid Mechanics 2022; 54: 443–467. DOI: 10.1146/annurev-fluid-030121-034728.

17.

Munnaf

Mouazen

. Removal of external influences from on-line vis-NIR spectra for predicting soil organic carbon using machine learning. Catena 2022; 211: 467. DOI: 10.1016/j.catena.2022.106015.

18.

Morgan

CLS

Ackerson

. VisNIR spectra of dried ground soils predict properties of soils scanned moist and intact. Geoderma 2014; 221-222: 61–69. DOI: 10.1016/j.geoderma.2014.01.011.